Kentino AI 192 Rome ArcProB70 TBD — 6× Intel Arc Pro B70 — EPYC Milan (Pré-commande)
Enterprise-grade platform, assembled and tested in the EU.
IN PREPARATION
Pre-order — Intel Arc Pro B70 shipping target Q3 2026
Kentino AI 192 Rome ArcProB70 TBD
192 GB VRAM Intel Xe2 Inference Server
6x Arc Pro B70 | EPYC Milan | TOPS TBD
Budget-oriented high-VRAM build targeting the Intel open-source inference stack. Pricing locked at Intel availability.
A 4U rack-mount inference server with six Intel Arc Pro B70 Creator cards (32 GB Xe2-HPG "Battlemage" each, 192 GB aggregate), one AMD EPYC 7643 Milan CPU (48C/96T), 384 GB DDR4 ECC, 2 TB NVMe boot, and a 2 kW ATX PSU (dual-PSU upgrade strongly recommended). Built for the Intel software ecosystem: OpenVINO 2025+, IPEX-LLM, llama.cpp SYCL backend, and vLLM-Intel forks. CUDA-only workloads do not run on this hardware.
Hardware
| Component | Detail |
|---|---|
| GPUs | 6x Intel Arc Pro B70 Creator 32 GB (Xe2-HPG "Battlemage", 250 W, PCIe 5.0 x16, dual-slot) |
| VRAM pool | 192 GB aggregate across 6 cards (no inter-card fabric — peer traffic over PCIe) |
| CPU | AMD EPYC 7643 Milan (48C/96T, 225 W, 128x PCIe 4.0 lanes) |
| Motherboard | ASRock Rack ROMED8-2T (SP3, 7x PCIe 4.0 x16, 8x DDR4 ECC, 2x 10 GbE, IPMI) |
| System RAM | 384 GB DDR4-2666 ECC RDIMM (6x 64 GB) |
| Boot / storage | 2 TB NVMe M.2 (PCIe 4.0 x4) |
| Power supply | 1x 2 kW ATX PSU (dual 2 kW synced upgrade strongly recommended) |
| Chassis | 4U rack-mount (6-slot layout) |
| Cooling | SP3 tower cooler (Arctic Freezer 4U-M) + front-to-back directed airflow (industrial fans) |
| Network | Onboard dual 10 GbE (Intel X550) |
Power envelope
- GPU draw: 6 x 250 W = 1 500 W (Intel-published TDP)
- System total at full load: ~1 825 W
- PSU total: 2 000 W (single) — only 8.75 % headroom
- Dual 2 kW synced strongly recommended — restores ~45 % headroom
Lane topology
ROMED8-2T provides 7x PCIe 4.0 x16 lanes. Six slots populated; one free for NIC upsell. Arc Pro B70 is PCIe Gen5 native; ROMED8-2T runs at Gen4 — bandwidth impact negligible for inference at 32 GB per card. No PCIe switch. No Xe-Link equivalent.
What you can run
All compatibility claims are Intel-software-stack paths (OpenVINO, IPEX-LLM, llama.cpp SYCL, vLLM-Intel). CUDA-only workloads do not run on this hardware. All figures cite published external sources and are subject to independent verification when cards ship.
LLMs — text / reasoning / coding
Chinese frontier
- Qwen3 / Qwen3.5 (Alibaba): Qwen3-235B-A22B Q4 (~132 GB) with long context headroom; Qwen3-Coder-480B-A35B Q2 (~160 GB); Qwen3.5-397B-A17B Q3 (~170 GB)
- GLM / Z.ai: GLM-4.5 / 4.6 / 4.7 Q4 (~177 GB) — fits with moderate KV
- Tencent Hunyuan: Hunyuan-Large Q3 (~160 GB); Hunyuan-A13B fp8 (~80 GB) if Xe2 fp8 path exposed in driver
- Others: Baidu ERNIE-4.5-424B Q3 (~180 GB); MiniMax-M1 Q3 (~180 GB); DeepSeek-R2 32B (6x concurrent streams)
Western frontier
- Meta Llama: Llama 3.3 70B Q6-Q8 with generous KV; Llama 4 Scout 109B/17B Q4 (~63 GB) comfortable
- Mistral: Mistral Small 3 / Magistral Small / Devstral Small 2 (24B) at bf16; Pixtral Large Q4-Q6
- OpenAI (open weights): gpt-oss-120b MXFP4 native (~80 GB) — if MXFP4 dequant available in Intel stack
- NVIDIA Nemotron: Llama-3.1-Nemotron Ultra 253B Q4 (~120 GB)
- Others: Gemma 3 27B bf16 multimodal; Phi-4 / Phi-4-reasoning 14B; Cohere Command R+ 104B Q4
Vision-Language Models
Qwen3-VL-8B / 32B; Qwen3-VL-30B-A3B MoE; InternVL3 up to 78B; InternVL3.5-38B; Llama 3.2 90B Vision Q4; Pixtral 12B; Molmo 72B Q4; Gemma 3 12B/27B multimodal; MiniCPM-V 2.6 / MiniCPM-o 2.6. Intel's OpenVINO has strong vision-tower support — VLM is a plausible day-one strength.
Image generation
FLUX.1 [dev] / [schnell] fp8 or Q4 GGUF via llama.cpp SYCL; SDXL / SD 3.5 Large via OpenVINO genAI runtime; HunyuanDiT; HunyuanImage-2.1 bf16 (~34 GB); Kolors 2.0; AuraFlow; OmniGen; PixArt-Sigma.
Video generation
Wan 2.2 T2V-A14B / I2V-A14B MoE (~54 GB bf16); Wan 2.2 TI2V-5B; HunyuanVideo 13B bf16; HunyuanVideo 1.5; CogVideoX-5B; Open-Sora 2.0; LTX-Video; Pyramid Flow; Mochi-1 Q4. Video is the weakest Intel path today — expect functional but not throughput-optimal at ship time.
Audio / Speech / TTS
- ASR: Whisper v3 large / turbo via OpenVINO (first-class Intel Whisper support); Parakeet-TDT; Canary; SenseVoice
- TTS: CosyVoice 2/3; Kokoro 82M; Stable Audio Open; XTTS v2; StyleTTS 2; Step-Audio-EditX
- Realtime / S2S: Kyutai Moshi; MusicGen / AudioGen / Bark; SeamlessM4T v2
Multi-model / multi-tenant serving
- 6 concurrent streams of a 32 GB Q4 model (one per card) — e.g. 6x Qwen3-32B Q4 agents
- Embedding-fleet at scale — 6x parallel BGE-M3 / E5 / Nomic Embed streams (OpenVINO-optimized)
- Mixed residency — 70B Q4 (tensor-parallel over 3 cards) + FLUX.1 (1 card) + Whisper-turbo (1 card) + Moshi (1 card)
Target workloads
- Intel-software evaluation pilot for CUDA-alternative LLM serving
- Embedding / reranker backend where VRAM-per-EUR dominates throughput requirements
- Budget Q4 frontier-MoE inference (Qwen3-235B, GLM-4.5/4.6/4.7) for small internal dev teams
- OpenVINO-native model deployment alongside existing Intel Xeon / Arc Pro pipelines
- VLM / OCR / document-processing backend (Intel's OpenVINO strength)
Measured performance
Intel-published specs | Subject to independent verification when cards ship
| Spec | Value |
|---|---|
| VRAM per card | 32 GB GDDR6 |
| Memory bandwidth class | ~450 GB/s per card |
| Xe Matrix Extensions (XMX) | Accelerated via OpenVINO / IPEX-LLM |
| fp8 path | Xe2 silicon — verify driver exposure at ship time |
No Kentino measured data. Intel-published specs subject to independent verification. Kentino will publish first-party tok/s / QPS / bandwidth numbers once the first unit passes burn-in.
Not ideal for
- CUDA-native workloads — no CUDA on Intel, expect migration friction
- Production SLA-critical deployments until Intel Arc Pro supply and tooling stabilize
- Frontier 600B+ MoE at Q4+ (requires 6x RTX Pro 6000 / 576 GB pool)
- Training workloads — Arc Pro is inference-first, framework maturity for distributed training is limited
- Customers who require measured benchmarks before purchase — this SKU is pre-order
Warranty and lead time
Kentino standard warranty (2 years parts, 1 year labor); Intel distribution terms supersede where stricter. Build includes assembly, BIOS configuration, driver install, burn-in testing, and functional verification. Reserve your first-wave delivery slot via the Kentino contact form. 30-day price-commit window at order.
Recommended add-ons
- Dual 2 kW synced PSU upgrade (single-PSU headroom is tight at 1 825 W draw — strongly recommended)
- Upgrade RAM to 512 GB DDR4 (2x 64 GB — two slots open)
- 4 TB NVMe secondary drive for model library
The questions buyers ask us most often before ordering a server.
How long does it take?
Machines built from components we hold ship quickly; anything requiring a specific GPU generation depends on supply. We give you a date before you pay, and if it moves we tell you rather than letting you find out.
Can the configuration be changed before you build it?
Almost always. GPUs, memory, storage and cooling are chosen per order, and the listed configuration is a starting point rather than a fixed package. If you need more VRAM, faster storage or a different cooling approach, say so before you order and we will quote the change.
Can I collect the server in person?
You can. Our warehouse is in Prague, and collection in person is welcome — most people who come use the visit to go through the machine with our engineer and ask the questions that are awkward over email. For orders within the Czech Republic we also try to deliver personally and walk you through the setup on site.
Can I talk to someone who actually understands the workload?
Yes. We have an engineer who works on AI systems specifically, not a general sales desk. If your question is about batch sizes, quantisation, interconnect or where your bottleneck will be, ask it — that conversation usually changes the configuration for the better.
Which model can I run on this configuration?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
How do you test a server before shipping?
We run it against actual AI workloads rather than synthetic scores: inference throughput, sustained load behaviour and thermals under continuous operation. You get the benchmark results with the machine, so the performance you were promised is the performance you can verify on day one.
Is the server ready to run when it arrives?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
Ships from our EU warehouse. Heavy items may require freight arrangement — contact us for a shipping quote and lead time. 2-year limited warranty with advanced RMA support; extended warranty available.
Not exactly what you need?
Tell us your workload and we'll spec this platform around it — GPUs, memory, storage and cooling matched to what you actually run.